Method for automatically planning trajectory for medical intervention

Abstract

To provide a method for automatically planning a trajectory to be followed during a medical intervention with a medical instrument targeting an objective anatomical structure of a patient.SOLUTION: A method for automatically planning a trajectory to be followed during a medical intervention with a medical instrument 120 targeting an objective anatomical structure 130 of a patient 110 includes the steps of: acquiring at least one medical image of the objective anatomical structure; determining a target point 145 on the previously acquired image; and generating a set of trajectory planning parameters from the medical image of the objective anatomical structure and the previously determined target point. The set of planning parameters includes coordinates of an entry point on the medical image. The set of parameters is generated using a machine learning method of neural network type. The present invention also relates to a guidance device 150 for implementing the obtained set of planning parameters.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] Technical field of the invention The field of the invention is that of aiding in the planning of medical interventions.

[0002] More specifically, the present invention relates to a method for automatically planning the trajectory of a medical instrument performed during a medical intervention, and to an associated guidance device.

[0003] The invention finds particular application in the context of medical interventions, where medical instruments are inserted into the anatomical structure of a subject, for example to cauterize a tumor in an organ, to perform a biopsy, to perform vertebroplasty or cementoplasty, or to stimulate an anatomical zone, which can optionally be assisted by a medical robot and / or by an augmented reality device. [Background technology]

[0004] Prior Art The prior art discloses techniques that make it possible to prepare medical interventions aimed at reaching target anatomical zones in a patient's target anatomical structure, such as the lungs, kidneys, liver, brain, tibia, knee, vertebrae, etc.

[0005] Traditionally, planning of medical interventions has been performed manually by an operator based on medical images obtained by conventional medical imaging techniques.

[0006] During planning, the operator defines a target point in the target anatomy and an entry point on the patient's skin adjacent to the target anatomy, and these two points define the linear trajectory of the medical instrument used during the medical intervention. Such an instrument may be, for example, a needle, a probe or an electrode.

[0007] The operator must pay attention to the trajectory followed by the medical instrument, since it must take into account many constraints necessary for the medical intervention to be carried out smoothly. For example, it may be important that the medical instrument does not pass through bones or blood vessels, especially those with a diameter of more than 3 millimeters, or pass through vital organs.

[0008] To assist the operator in selecting approach points according to the target point, a planning technique has been developed in which one or more approach points are automatically suggested to the operator depending on previously defined constraints by associating a score with each corresponding trajectory according to predefined criteria.

[0009] This type of technique is described, for example, in the patent application published under the number US Patent Application Publication No. 2017 / 0148213A1 entitled "Planning, navigation and simulation systems and methods for minimally invasive therapy". The method described in said patent application determines the trajectory using conventional image processing algorithms in which the image is segmented so that the constraints related to the trajectory can be minimized. For example, during surgery on the brain, the trajectory is determined by optimization of several parameters such as the number of fibers affected, the distance between the border of the cortical sulcus and the target, minimizing the volume of white and / or gray matter displaced by the trajectory, etc. Summary of the Invention [Problem to be solved by the invention]

[0010] However, the main disadvantage of the techniques in the prior art is that they are generally based on the minimization of constraints selected by the operator to create a theoretical model, which is often insufficient and incomplete. In addition, they require a systematic segmentation of the image in order to be able to optimally calculate the different possible trajectories. This segmentation sometimes proves to be inaccurate and incomplete, which may lead to errors in the trajectories used by the medical instrument.

[0011] Furthermore, these techniques do not take into account possible deformation of a medical instrument, such as a needle, as its end is inserted into a patient's body.

[0012] Finally, a skilled operator also periodically intervenes to select from the images those areas that should be avoided, such as blood vessels, and those through which the medical instrument must pass, in order to determine the optimal trajectory of the medical instrument.

[0013] Operator intervention has proven to be both cumbersome and restrictive, as such intervention requires a great deal of care and experience on the part of the operator in this type of intervention.

[0014] None of the current systems are able to simultaneously meet all of the required needs, i.e., to make available improved techniques for automatically planning medical interventions aimed at reaching targets in the patient's target anatomy, independent of the operator, while allowing more accurate and more reliable planning. [Means for solving the problem]

[0015] Disclosure of the Invention The present invention aims to overcome all or some of the disadvantages of the prior art discussed above.

[0016] To this end, the invention relates to a method for automatically planning a trajectory to be followed during a medical intervention with a medical instrument targeting a target anatomical structure of a patient, said automatic planning method comprising the steps of: - acquiring at least one medical image of an anatomical structure of a subject; - determining a target point in a previously acquired image; - generating a set of trajectory planning parameters based on an image of the target anatomical structure and on previously determined target points, the set of planning parameters including coordinates of an entry point in the medical image; Includes.

[0017] Such methods are used prior to a medical intervention and make it possible to provide a set of parameters that guide the doctor or surgeon during the manipulation of a medical instrument, which may be a needle, a probe, an electrode or any other medical instrument that can be inserted into the patient's body, using reference points associated with the patient, which are generally three-dimensional in order to guide the medical instrument in space.

[0018] The aim of a medical intervention is to reach a target anatomical zone of a patient's body, for example to cauterize a tumor in an organ, to perform a biopsy, to perform vertebroplasty or bone cement therapy, or to stimulate an anatomical zone. The target anatomical zone is located in or on a target anatomical structure of the patient. Such target anatomical structures are, for example, the lung, kidney, liver, tibia, knee, vertebrae, or brain.

[0019] The medical images used for planning have been obtained, for example, by computed tomography, by magnetic resonance imaging, by ultrasound, by positron emission tomography, or by any other medical imaging method.

[0020] According to the invention, the set of parameters is generated by implementing a neural network type machine learning method previously trained on what are referred to as a set of medical training images, each medical training image containing a target anatomical structure similar to the target anatomical structure of the patient, each medical training image being associated with the coordinates of predetermined target points and at least one entry point.

[0021] Therefore, the planning method can be used by any operator, who only needs to select target points in a medical image.

[0022] It should be noted that the present planning method is based on machine learning of similar medical images, each of which has associated entry and target points.

[0023] Similar medical images are understood to mean images obtained by the same or equivalent imaging method and containing the same target anatomical structure in the medical images taken for any individual. It should be noted that the type of medical intervention, the type of medical instrument, or the target anatomical structure targeted can be different without compromising the accuracy of the obtained planning parameters. Learning actually makes it possible to analyze new images in order to determine the optimal trajectory to the target point selected by the operator in the medical image of the target anatomical structure of the patient.

[0024] It should be noted that the medical training images are generally associated with the entry points that will actually be used during the medical intervention that an individual undergoes, and the target points that the instrument will actually reach following its insertion. To complete the set of medical training images, medical images associated with the envisaged entry points selected by the operator can be added to the set.

[0025] In addition, the automatic planning method is advantageously based on learning non-segmented medical images, i.e. images where all or part of the image is characterized according to the type of tissue, organ or vessel present in that part of the image, and therefore the processing of images by the planning method is faster.

[0026] A set of medical training images is generally contained in a database or in a medical image bank.

[0027] The automatic planning method generally provides planning parameters for at least one possible trajectory. If the automatic planning method provides planning parameters for several possible trajectories, the operator usually manually selects the trajectory that he or she considers to be the best. It should be noted that a trajectory is generally considered to be the best when it meets a number of criteria specific to the medical intervention, such as the angle of incidence to tissue interfaces (e.g., skin, liver capsule, etc.), the proximity of blood vessels, organs or bone structures in the trajectory, etc.

[0028] It should be noted that the automated planning method is performed prior to any medical, surgical or therapeutic procedure.

[0029] In certain embodiments of the present invention, a machine learning method determines the coordinates of the entry point based on the acquired medical image and a previously determined target point in the acquired medical image.

[0030] In a particular embodiment of the present invention, the machine learning method first generates for each pixel or voxel of a medical image acquired in two or three dimensions, respectively, a probability of being an intrusion point, the coordinates of the intrusion point corresponding to the coordinates of the pixel or voxel with the maximum probability.

[0031] In certain embodiments of the present invention, the set of similar medical images includes multiple identical images, each identical image being associated with a different entry point.

[0032] Thus, the set of medical images includes possible trajectory deformations for the medical instrument, improving learning.

[0033] Advantageously, the set of similar medical images comprises a number of identical images, each identical image being associated with a different entry point selected by a different operator.

[0034] Therefore, the resulting planning parameters are more accurate since they are less sensitive to the choice of a particular operator. It should be noted that the accuracy of the resulting planning parameters depends on the number of operators involved in the analysis of the same medical image during the learning phase.

[0035] Preferably, the set of similar medical images includes at least three identical images, each identical image associated with a different entry point by a different operator.

[0036] Therefore, at least three operators are involved in the generation of the database containing the set of medical images used during the learning phase.

[0037] In certain embodiments of the present invention, information related to a target anatomy is associated with each medical image in the set of medical images, the information including the type of target anatomy or tumor present in the target anatomy, and the machine learning method is trained on a plurality of medical image sets restricted to images related to the same type of anatomy or tumor.

[0038] In certain embodiments of the present invention, the automated planning method also includes a step of assigning a score to a trajectory defined between an entry point of the set of planning parameters and a predetermined target point in the acquired images.

[0039] Thus, the operator is assisted in his / her own selection of a trajectory from among the possible trajectories provided by the automated planning method. Scores are generally assigned according to criteria that are specific to the medical intervention.

[0040] The trajectory defined between the entry point of the set of planning parameters and the target point previously determined in the acquired images is generally linear. However, to take into account the stiffness of the medical instrument, it can be assumed that the trajectory is curved, for example substantially following the arc of a circle having the maximum radius of curvature. Generally, a curved trajectory is either concave or convex. In other words, the derivative of the curved trajectory generally has a constant sign, negative or positive, between the entry point and the target point.

[0041] Preferably, the assignment of trajectory scores is based on the following criteria: - proximity of blood vessels, - Proximity of organs, - Proximity of bone structures, - the angle of incidence relative to the tissue interface, - length of the orbit, -Weakness of the organisations through which the trajectory passes It depends on at least one of the following:

[0042] In certain embodiments of the invention, the assignment of a trajectory score takes into account the probability that the medical device will deform upon contact with a tissue interface.

[0043] This deformation generally occurs when the medical device has flexible portions, ie portions that can deform when contacted with a tissue interface during insertion of the medical device through the skin of a patient.

[0044] In certain embodiments of the invention, the assignment of a trajectory score takes into account relapse rates or recovery times associated with trajectories similar to the planned trajectory.

[0045] Thus, if a planned trajectory results in too high a relapse rate or too long a recovery time for the patient, the score assigned to the trajectory will be negatively affected.

[0046] In certain embodiments of the present invention, the automated planning method also includes a step in which the score assigned to the trajectory is compared to a threshold score, and the trajectory is validated if the trajectory score is equal to or greater than the threshold score.

[0047] In certain embodiments of the present invention, the automated planning method also includes modifying the approach point if the score assigned to the trajectory falls below a threshold score.

[0048] In certain embodiments of the invention, the acquired medical images are two-dimensional or three-dimensional.

[0049] In certain embodiments of the invention, the medical images are obtained by magnetic resonance, by ultrasound, by computed tomography, or by positron emission tomography.

[0050] The invention also relates to a device for guiding a medical instrument, comprising means for guiding the medical instrument according to a set of planning parameters obtained by an automatic planning method according to any one of the preceding embodiments.

[0051] The guidance device may be a robotic device, a navigation system associated or not with a robotic device, an augmented reality device, a patient-specific guide, or a three-dimensional model of the patient's anatomy.

[0052] It should be noted that the device for guiding the medical instrument makes it possible to accompany a doctor performing a medical intervention.

[0053] BRIEF DESCRIPTION OF THE DRAWINGS Other advantages, objects and particular features of the present invention will become apparent from the following non-limiting description of at least one particular embodiment of the device and method that are the subject of the invention, with reference to the accompanying drawings, in which: [Brief description of the drawings]

[0054] [Figure 1] FIG. 1 is a schematic diagram of a medical intervention in which a medical instrument is guided according to a set of parameters established by the automatic planning method according to the invention. [Diagram 2] FIG. 1 is a block diagram of an automated planning method in accordance with certain embodiments of the present invention. [Diagram 3]3 is an example of a medical image acquired during a first step of the planning method of FIG. 2. [Figure 4] 3 is an example of a medical image used during training of a neural network performed by the method of FIG. 2. [Diagram 5] FIG. 3 is a schematic diagram of the neural network training stage performed by the method of FIG. 2. [Figure 6] FIG. 6 is a schematic diagram of the evolution of a neural network carried out by the method of FIG. 2 and trained according to the training steps of FIG. 5. [Figure 7] FIG. 3 is a schematic diagram of the evolution of a neural network trained according to the method of FIG. 2 and in accordance with alternative training stages. [Figure 8] The figure shows two medical images of the same patient, one with a medical instrument inserted and the other corresponding to the same view without the instrument, said images being used while training a neural network configured to define a curvilinear trajectory of the medical instrument. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0055] Detailed Description of the Invention This description is given without limitation, and each feature of an embodiment can be advantageously combined with any other feature of any other embodiment.

[0056] It should be noted here that the figures are not to scale.

[0057] Examples of Specific Embodiments FIG. 1 is a schematic diagram of a medical intervention in which a patient 110 lying on a table 115 is treated employing a medical instrument 120. In this non-limiting example of the invention, the medical intervention corresponds to the ablation of a tumor in a target anatomical structure 130, here the liver of the patient 110, by the medical instrument 120, in this case a semi-rigid needle. The medical intervention here is a percutaneous procedure in which the body of the patient 110 is not incised. In addition, the medical intervention can be performed according to different treatment parameters. Such treatment parameters are, for example, the duration and power of an ablation treatment, the voltage applied in the case of treatment by electroporation, or the frequency applied in the case of treatment by radiofrequency. It should be noted that this example is given by way of illustration and that a person skilled in the art can implement the invention described below for any type of medical intervention using any medical instrument targeting a target anatomical structure of a patient.

[0058] The medical tool 120 in this example is advantageously guided by the device 150 along a linear path, with prior establishment of a set of planning parameters, including the coordinates of an entry point 140 at the skin level of the patient 110, or even the angle to be followed in a three-dimensional reference system associated with the patient 110, to target a pre-determined target point 145. The set of planning parameters is established by an automatic planning method 200 according to the invention, as shown in block diagram form in FIG.

[0059] The method 200 for automatically planning a trajectory to be followed by a medical instrument 120 during a medical intervention comprises a first step 210 of acquiring at least one medical image of a target anatomical structure 130 of a patient 110 .

[0060] Medical images are generally taken prior to a medical intervention using equipment dedicated to medical imaging, such as a Magnetic Resonance Imaging (MRI) machine, a CT scanner, a spectral scanner or an ultrasound machine.

[0061] 3 presents an example of a medical image 300, showing a model, commonly referred to as a phantom, obtained by computed tomography and corresponding to a target anatomical structure 130 of a patient 110. The medical image 300 corresponds to a cross-section of the patient 110 through a plane substantially perpendicular to the axis of the spinal column of the patient 110. In addition to the target anatomical structure 130, the medical image 300 notably reveals the vertebrae 310 of the spinal column and six ribs 320.

[0062] In the previously acquired medical image 300, the target point 145 is determined during a second step 220 of the automated planning method 200, either manually by an operator or automatically by image analysis.

[0063] The target point 145 is associated with coordinates in the medical image 300. These coordinates may be two-dimensional or three-dimensional depending on the type of medical image acquired. In the case of a two-dimensional medical image 300, the target point 145 substantially corresponds to a pixel of the image. In the case of a three-dimensional medical image 300, the target point 145 substantially corresponds to a voxel of the image.

[0064] A machine learning algorithm, here of neural network type, is loaded during the third step 230 of the automatic planning method 200 to determine, from the medical image 300 and from the target point 145, the coordinates of an entry point of a set of parameters for planning the trajectory to be followed by the medical tool 120.

[0065] The neural network is pre-trained during a learning phase 290 on a set of medical training images, each of which includes a target anatomy similar to the target anatomy 130. The medical training images have generally been acquired for individual cohorts, and each medical training image is generally associated with coordinates of target and entry points that have been previously determined by at least one operator.

[0066] Advantageously, the set of medical training images includes several times the same medical image but associated with different entry points as determined by typically at least three operators.

[0067] 4 shows an example of the same medical image 400, each time containing the same target point 420. This medical image 400 was included in a set of medical training images nine times and was processed by three separate operators 01, 02 and 03, who each entered three entry points, 410, respectively. 01 , 410 02 , 410 03 is offered.

[0068] Training of the neural network can be advantageously restricted to images associated with a given item of information, such as an anatomical region of interest, or a type of tumor present in the anatomical structure of interest, in order to reduce variability in the set of planning parameters that the neural network can obtain, thereby improving consistency.

[0069] It should be noted that there may be hardware limitations to training the neural network, especially when the set of medical training images includes three-dimensional images of the target anatomical structure. To overcome these hardware limitations, it is possible to reduce the resolution of each medical image, but this comes at the risk of reducing the accuracy of the parameters obtained by the neural network. It is also possible to limit training to trajectories parallel to a given plane, such as a plane perpendicular to the axis of the patient's spine. Another solution to overcome hardware limitations may be to use chips, commonly referred to as tensor processor units, that are dedicated to machine learning.

[0070] As shown in more detail in Fig. 5, the stage 290 of training the neural network generally comprises two main steps 510, 520 that can be repeated and requires a database 501 containing a set of medical images, each image associated with an entry point and a target point. Optionally, information about the characteristics of the instrument used to perform the intervention, such as the instrument's length or its stiffness coefficient, is also associated with each medical image in the database 501. After the training stage 290, a possible testing stage 550 can be performed.

[0071] A database 501 of medical images is divided into three databases 502, 503, 504 containing different medical images. The three databases 502, 503, 504 are referred to as the training base, the validation base, and the test base, respectively.

[0072] In this non-limiting example of the invention, 60-98% of the medical images in database 501 are grouped into a training base 502, 1-20% into a validation base 503, and 1-20% into a test base 504. The percentages are generally a function of the number of images in database 501 and are given here for illustration purposes.

[0073] During a first step 510 of the training phase, medical images 515 of the training base 502 are used to determine the weights W and biases b for each neuron of the neural network 530 used to obtain the coordinates of the entry points of the trajectory planning parameter set.

[0074] To determine the weights W and biases b of each neuron, each medical image 515 of the training base 502 is transformed into two images, i.e., the target point C e The first transformation 5151, which includes only the target point C e and a second deformation 5152, which includes both the first deformation 5151 of the medical image and the predetermined entry point p. The neural network 530 then makes a prediction 535 regarding the location of the entry point p' from the first deformation 5151 of the medical image. The coordinates of the predicted entry point p' are compared with the coordinates of the location of the predetermined entry point p associated with the second deformation 5152 of the medical image. The error between the coordinates of the predicted entry point p' and the coordinates of the predetermined entry point p is then used to adjust the parameters W and b of each neuron of the neural network 530. At the end of the first step 510 of the training phase, a model 518 is obtained.

[0075] During a second step 520 of the training phase, the weights W and biases b of each neuron of the neural network 530 are validated using medical images 525 of the validation base 503 , advantageously different from the medical images 515 .

[0076] During this second step 520 of the training phase 290, the target point c v A deformation 5251 of each medical image, which includes only the positions of the intrusion points d', is proposed to the neural network 530. The neural network 530 then makes a prediction 536 regarding the position of the intrusion point d'. The coordinates of the predicted intrusion point d' are compared with the coordinates of the position of the predefined intrusion point d associated with the medical image 525 used for validation. The error between the coordinates of the predicted intrusion point d' and the coordinates of the predefined intrusion point d is then used to validate the parameters W and b of each neuron of the neural network 530 determined in the first step 510.

[0077] At the end of this second step 520, if the prediction error of the neural network is too large, the neural network 530 is retrained according to the two steps 510 and 520 of the training phase 290 described above, by reusing the same medical training images 515 and validation images 525.

[0078] Alternatively, during retraining of the neural network 530, the first step 510 uses all or a portion of the validation images 525. The second step 520 of retraining the neural network uses the same number of training images 515 as the number of validation images 525 used in the first step 510 of retraining.

[0079] It should be noted that the neural network 530 can be retrained as many times as necessary to reduce the prediction error.

[0080] Once the two steps 510, 520 of the training phase 290 have been carried out at least once, the final performance of the neural network can be tested during a possible testing phase 550 with medical images 555 of the test base 504. These medical images 555, advantageously different from the images 515 and 525, make it possible to verify that the neural network 530, configured with the parameters W and b for each neuron, is able to predict with good accuracy the coordinates of the entry point in all situations likely to be encountered by the neural network 530. A comparison is therefore made between the coordinates of the entry point f' as predicted by the neural network 530 and a given entry point f in a so-called test medical image 555. This comparison is identical to the one carried out during the second step 520 of the training phase. However, in contrast to step 520, this testing phase 550 does not result in a new training cycle of the neural network 530. If at the end of step 550 the performance of the neural network 530 is not good, the training phase 290 is restarted with a new untrained neural network.

[0081] The images 555 used in the testing phase 550 are generally images of target points c in the subject's anatomical structure to optimally test the predictive capabilities of the training network 530. t It should be noted that the sequences are carefully selected to encompass a variety of positions.

[0082] In an alternative training phase, the neural network can be trained to provide for each pixel or voxel of the medical image a probability that it actually corresponds to an entry point. The set of medical images used for this alternative training can be identical to the set of medical images used previously. However, for this alternative training, it may be preferable to use medical images with several entry points in the same image. Advantageously, the entry points displayed in the same image are determined by at least three different operators. The alternative training of the neural network is performed in three steps similar to the training phase described above.

[0083] The previously trained neural network makes it possible, during a fourth step 240 of the automatic planning method 200, to determine, based on the analysis, at least a set of parameters for planning the trajectory to be followed by the medical instrument 120.

[0084] When the neural network is trained according to the training phase 290, the neural network 530 provides the three-dimensional coordinates (x, y, z) of the entry point in the acquired medical image from the medical image I and from the coordinates of the target point T, as shown in FIG. 6.

[0085] When the neural network is trained according to the alternative training phase, the neural network 530 provides, from the medical image I and from the coordinates of the target point T, for each pixel or voxel of the medical image, a probability of being an intrusion point, as shown in Figure 7. The pixel or voxel with the highest probability is then selected as being the intrusion point.

[0086] 2 includes a fifth step 250, which is performed when a trajectory is determined using a set of planning parameters generated by a neural network. During this fifth step 250, a score is assigned to the trajectory defined by the straight line connecting the entry point and the target point.

[0087] For example, the score assigned to a trajectory can range from 0 to 100, with a score of 100 corresponding to an ideal trajectory score.

[0088] In a variant of this particular embodiment of the invention, the trajectory is curvilinear and is obtained, for example, by calculating a most likely trajectory on previously segmented acquired medical images or by a neural network previously trained on trajectories followed during previous medical interventions by similar or identical medical instruments, in particular with regard to stiffness and length, and the set of parameters then includes additional parameters making it possible to define a predicted trajectory between the entry point and the target point.

[0089] To illustrate these alternative embodiments of the invention, Fig. 8 shows two medical images 810, 820 of a patient 830 with and without a medical tool 840. By differencing the two medical images 810 and 820, it is possible to determine the trajectory actually followed by the medical tool 840. This trajectory can also be determined by performing a recognition of the medical tool 840 in the medical image 810, for example by detecting strong variations in intensity or contrast in the pixels / voxels of the medical image 810, in order to define the path of the medical tool 840 in the medical image 810.

[0090] The trajectory score is generally determined based on criteria that can be ranked in order of importance. It should be noted that the example criteria set forth below are not limiting and other criteria specific to a given medical intervention can be used to determine the trajectory score.

[0091] The trajectory score can be calculated, for example, as a function of the trajectory's proximity to blood vessels, since there is a high probability that the trajectory of the medical instrument will pass through a blood vessel, which will increase the risk of bleeding. Thus, the more blood vessels present in a trajectory, the lower the score assigned to that trajectory.

[0092] It should be noted that in the assessment of this score, the size of the blood vessels can be taken into account. For example, if blood vessels with a diameter of 3 mm or more are located on or near the trajectory calculated by the neural network, a point is automatically subtracted from the score, for example 50 points on a scale of 0 to 100, since these blood vessels may be important for the patient. If the passing blood vessel turns out to be the vena cava, portal vein or aorta, the score is automatically equal to 0, which may be the case especially when removing a tumor from the liver.

[0093] A trajectory score may also be calculated according to the proximity of the trajectory to organs and / or bone structures.

[0094] In fact, for some interventions, e.g. soft tissue interventions, no bony structures should be located in the trajectory, otherwise the score assigned to that trajectory is zero.

[0095] In other interventions, e.g. interventions on bony structures such as the knee or shoulder, passing through a bony structure does not negatively impact the assigned score, more precisely, if the trajectory passes through a given bony structure, the assigned score may be increased.

[0096] With regard to organs, the orbit score is generally decreased if organs at risk, such as lungs, intestines or muscles, are located at least close to the orbit, as is the case for nerves, bile ducts, ligaments, tendons or adjacent organs of the anatomical structure of interest, located at least close to the orbit.

[0097] The trajectory score can also be calculated according to the angle of incidence to the tissue interface at the entry point of the trajectory.

[0098] For example, when inserting a semi-rigid needle along a trajectory tangential to a tissue interface such as the skin or liver capsule, there is a risk that the needle will bend and not follow the planned trajectory. The smaller the angle between the trajectory and the tissue interface, the lower the trajectory score. This criterion may reflect the fact that an optimal trajectory corresponds to an angle between the tissue interface and the trajectory of more than 20°.

[0099] The trajectory score can also be calculated by the angle of incidence of the trajectory onto the bone structure.

[0100] For example, in the case of interventions on bony structures, there is a risk of the medical instrument slipping on the bone when inserted tangentially to the bone, so the criteria reflect the fact that the greater the angle between the trajectory and the bony structure, the lower the trajectory score.

[0101] A trajectory score can also be calculated according to the length of the trajectory to minimize the length of the trajectory and the inherent risk of causing damage inside the patient's body.

[0102] A trajectory score can also be calculated according to the vulnerability of the tissues it passes through.

[0103] For example, in the specific case of an intervention on a patient's brain, the trajectory score may be reduced if the planned trajectory passes through fragile tissue.

[0104] In case of insertion of semi-rigid needles, a score can also be calculated according to the probability of the needle being deformed during insertion, this probability being calculated based on information about the type of needle used, such as its length, its stiffness coefficient or the shape of the needle bevel, in combination with previously determined information, i.e. the type of tissue passed, the angle of incidence and / or the length of the trajectory.

[0105] It should be noted that in order to calculate the trajectory score, the acquired medical images may have been pre-segmented to identify different types of elements present in the acquired images, such as tissues, blood vessels, bone structures, etc., that are located on or in close proximity to the trajectory defined between the predicted entry point and the predefined target point. This segmentation of the acquired images is only used when the trajectory of the medical instrument is generated, and not during the generation of the trajectory by the neural network.

[0106] Trajectory scores obtained according to criteria specific to a medical intervention can be weighted as a function of relapse rate and / or by recovery time.

[0107] Regarding relapse rate, the score obtained is reduced if the trajectory planned by the neural network is similar to a trajectory used with the same treatment parameters during a previous medical intervention with the same medical device, and in which the relapse rate of individuals who underwent these medical interventions was significant.

[0108] Similarly, with respect to recovery time, if prior observed recovery times for individuals undergoing medical interventions on a trajectory similar to the planned trajectory are long, for example greater than 3 days, the resulting score will be reduced.

[0109] The score assigned to the planned trajectory is then compared to a threshold score during a sixth step of the automated planning method 200.

[0110] For example, on a scale of 0 to 100, a planned trajectory may be validated only if the assigned score is 50 or greater. Preferably, a planned trajectory is validated if its score is 70 or greater.

[0111] If the score assigned to the trajectory falls below the threshold score, the operator has the option to manually modify the approach points during a possible seventh step 270 of the automated planning method 200. This modification is done, for example, via a graphical interface until the score of the modified trajectory exceeds the threshold score.

[0112] Alternatively, the trajectory can be automatically modified using a gradient algorithm, a graph algorithm, or any other optimization algorithm (Momentum, Nesterov Momentum, AdaGrad, RMSProp, Adam, etc.).

[0113] Finally, when the score for the trajectory provided by the possibly modified neural network is greater than or equal to the threshold score, the trajectory is validated during an eighth step 280 of the automated planning method 200.

[0114] The verified trajectory can then be used during the medical intervention to guide the insertion of the medical instrument 120 into the target anatomical structure 130 of the patient 110 with very good accuracy such that the medical intervention is most likely to be successful.

[0115] It should be noted that the reference frame used for navigation generally corresponds to the table 115 on which the patient 110 lies. The coordinates of the target points are advantageously transformed into a navigation reference frame to which the feature points of the patient 110 have been pre-calibrated. This navigation reference frame transformation and calibration operation is common.

[0116] The guidance device 150 can then be used to guide the medical instrument 120 by following the set of planning parameters of the verified trajectory.

[0117] The guidance device 150 may be a robotic device, a navigation system associated or not with a robotic device, an augmented reality device, a patient-specific guide 110, or a three-dimensional model of the patient's 110 anatomy.

[0118] The augmented reality device can be, for example, glasses, where the planned trajectory is projected onto at least one of the lenses of the glasses. The augmented reality device can also be a screen placed close to the patient 110, where the screen displays the planned trajectory. The augmented reality device can also include a projector that projects the planned trajectory onto the body of the patient 110, or can be a holographic device.

[0119] The guidance device may include optical navigation means, electromagnetic navigation means, or an inertial unit having acceleration and rotation sensors.

[0120] A set of planning parameters of the verified trajectory can be used to construct a patient-specific guide 110. This specific guide is generally used in the context of medical interventions, such as open surgery on bone structures. It should be noted that the patient-specific guide is generally a 3D printed, personalized, single-use medical device. The patient-specific guide helps to prevent inaccuracies during surgery in order to perform the intervention as planned. The specific guide generally conforms to the shape of the bone structure corresponding to the target anatomy and can allow for guiding the insertion of medical instruments according to the orientation of the planned trajectory.

[0121] For some medical interventions, a three-dimensional model of the anatomy of the patient 110, commonly referred to as a phantom, is constructed to enable training prior to the medical intervention, and the three-dimensional model of the anatomy of the patient 110 can advantageously include a representation of the entry points of the medical instrument as defined in the set of planning parameters of the trajectory obtained by the automatic planning method 200.

[0122] The results obtained by the automatic planning method 200 can also be used to present a medical intervention to a professional, such as a doctor, surgeon or radiologist, or even to a patient 110 undergoing the medical intervention. These results can also be used to train a professional in the medical intervention.

Claims

1. A method (200) for automatically planning a trajectory to be followed during a medical intervention with a medical instrument (120) targeting a target anatomical structure (130) of a patient (110), comprising: acquiring (210) at least one medical image (300) of the anatomical structure of the subject; determining (220) a target point (145) in a previously acquired image (300); generating (240) a set of trajectory planning parameters from the medical image of the target anatomical structure and from the previously determined target points, the set of planning parameters including coordinates of an entry point (140) in the medical image (300); Including, 1. The method of automated planning, characterized in that the set of parameters is generated using a machine learning method (530) of neural network type, previously trained on a set of so-called medical training images, each medical training image comprising a target anatomical structure similar to the target anatomical structure (130) of the patient (110), and each medical training image is associated with previously determined coordinates of a target point and at least one entry point.

2. The automated planning method of claim 1 , wherein the machine learning method determines coordinates of the entry point from the acquired medical image and from the target point previously determined in the acquired medical image.

3. 3. The automated planning method according to claim 1 or 2, wherein the machine learning method first generates for each pixel or voxel of the medical image acquired in two or three dimensions, respectively, a probability of being an entry point, the coordinates of the entry point corresponding to the coordinates of the pixel or voxel with the maximum probability.

4. The automated planning method according to any one of claims 1 to 3, wherein the set of similar medical images comprises a number of identical images, each identical image being associated with a different entry point.

5. The automated planning method according to any one of claims 1 to 4, wherein the set of similar medical images comprises a number of identical images, each identical image being associated with a different entry point selected by a different operator.

6. 6. The automated planning method according to claim 1 , wherein information related to the target anatomy is associated with each medical image of the set of medical images, the information comprising a type of the target anatomy or a tumor present in the target anatomy, and the machine learning method is trained on a plurality of the medical image sets restricted to images related to the same type of anatomy or tumor.

7. The automated planning method according to any one of claims 1 to 6, also comprising the step of assigning a score to a trajectory defined between the entry point of the set of planning parameters and the previously determined target point in the acquired images.

8. The acquired images are mapped and the assignment of the trajectory score is determined based on the following criteria: Proximity of blood vessels, Proximity of organs, Proximity of bone structures, the angle of incidence relative to the tissue interface, the length of the track; The vulnerability of the tissue through which the trajectory passes 8. The automated planning method of claim 7, wherein the time is a function of at least one of:

9. 9. The automated planning method of claim 7 or 8, wherein the assignment of the trajectory score takes into account the probability that the medical device will deform when in contact with a tissue interface.

10. The automated planning method of any one of claims 7 to 9, wherein the assignment of the trajectory score takes into account recurrence rates associated with trajectories similar to the planned trajectory.

11. The automated planning method of any one of claims 7 to 10, wherein the assignment of the trajectory score takes into account recovery times associated with trajectories similar to the planned trajectory.

12. The automated planning method of any one of claims 7 to 11, also comprising a step of comparing the score assigned to the trajectory with a threshold score, and wherein the trajectory is validated if the trajectory score is equal to or greater than the threshold score.

13. The automated planning method of any one of claims 7 to 12, also comprising the step of modifying the entry point if the score assigned to the trajectory falls below the threshold score.

14. A device for guiding a medical instrument, said device comprising means for guiding said medical instrument according to a set of planning parameters obtained by the automatic planning method according to any one of claims 1 to 13.

15. 15. The navigating device of claim 14, which is either a robotic guidance device, a navigation system with or without associated with a robotic device, an augmented reality device, a patient-specific guide, or a three-dimensional model of the anatomy of the patient.

Images